HYDRA
Hybrid Deep-learning for Residual Analysis
A compact 1-million parameter hybrid GRU-Transformer model that improves National Water Model streamflow predictions by up to 48% across unregulated Appalachian watersheds.
The Problem
NOAA's National Water Model (NWM) provides real-time streamflow forecasts across the continental United States, but exhibits systematic errors in headwater catchments where complex terrain and heterogeneous land cover challenge physics-based approaches.
Events like Hurricane Helene (September 2024) underscored the critical need for accurate streamflow predictions in southern Appalachian watersheds, where NWM errors can exceed 50% during peak flows — precisely when accuracy matters most.
Study Region
Study Region
Southern Appalachian headwaters in NC, VA, and TN
Three unregulated USGS gauging stations in the southern Appalachian highlands, spanning the New River and Watauga River basins in Virginia and North Carolina.
Mixed deciduous-coniferous forest at 500–1400 m elevation. Humid subtropical climate with orographic precipitation enhancement. Study period: 2010–2020 (hourly).
Southern Appalachian Biome
The study region spans the Blue Ridge physiographic province, characterized by temperate deciduous forests, steep terrain, and high annual precipitation (1,200-2,000 mm). This creates flashy, responsive watersheds where streamflow can change rapidly during storm events.
Hydrometeorological Regime
The region experiences orographic enhancement of precipitation, with the Blue Ridge escarpment forcing moist air upward. Tropical remnants and atmospheric rivers can produce extreme rainfall, while baseflow is sustained by fractured bedrock aquifers.
Research Motivation: Hurricane Helene
Hurricane Helene (September 2024) devastated this region, with catastrophic flooding in western North Carolina causing over 200 deaths and billions in damages. NWM forecasts significantly underestimated peak flows during this event. This research aims to improve streamflow predictions in mountainous terrain where operational models struggle most, potentially enabling better early warnings for future extreme events.
How to Use This Tool
Understand the Model
Learn how HYDRA combines GRU temporal encoding with Transformer attention to correct NWM errors in real-time.
Explore Experiments
Compare 19 configurations across 3 sites with interactive hydrographs, error distributions, and performance metrics.
Review Evaluation
Examine skill scores with bootstrap confidence intervals, significance tests, and flow regime analysis.